WTW

Consistent with the two nearest peers in this pocket, both also at B. The modelling engine is what actuaries license this for: it fits generalised linear models, gradient boosting machines and classification models, and the surrounding environment exists to build, compare and deploy them.

What holds it off A is the same residue as the peers and the vendor states it plainly: the rate delivery component is described as a full rating engine replacement, and a rating engine executes a published rate table deterministically. Strip every model and a production rating engine handling billions of quotes a day remains, which insurers buy in its own right. Membership settled on the services hybrid rule: this is a broking and consulting group that also ships named, separately licensed products, so it is graded on the products.

Asserts governance exists without describing it, which is the settled B, though the vocabulary is better chosen than most. The published language is well governed predictive modelling, automation without compromising business goals, and governance requirements met when deploying open source code.

None of it states who approves a rate change before it reaches the market, what an optimiser may move a price by without sign off, or what review applies to a model built inside the environment by a business user. For a product whose output is a live rate served to consumers at scale, the absence of a published approval boundary is the material gap.

Named mechanisms rather than asserted properties, which is what separates this from a C. Interpretable machine learning is stated as a design commitment with the explicit goal that a user can explain their models and their data. Governance requirements are stated for deploying open source code into production.

Most concretely, the environment retains historic models so that policy adjustments can be priced on the model version in force at the time and regulatory requirements met, which is model versioning described as a working feature rather than as a policy.

Held at B and well short of the A held by the nearest peer in this pocket: no validation methodology, no accuracy or discriminatory power figures, no drift monitoring, no external assessment, and nothing at all published about the generative component added in the current release.

Meets the B bar and stops there, which is notable for a vendor of this age and installed base. Manitoba Public Insurance is named as adopting the modelling and pricing products for generalised linear model ratemaking, with the engagement described in some detail, and a senior director of the vendor is quoted about it. The claim that most of the world's leading insurance groups use the software is the aggregate assertion this index sees constantly and credits nowhere.

No named institution carries a quantified outcome, no customer executive is quoted, and no current independent analyst placement surfaced. Thirty years in market and one named reference located is a thin public record.

Nothing published on training data, retention or whether anything is computed across the client base. The question has weight here because the vendor sits at the centre of an industry as both broker and software supplier, and because the same group advises insurers competing with each other in the same markets. Neither claimed nor denied.

Nothing published. The gap is worth stating precisely for a hosted pricing environment: the data being modelled is policyholder and quote level, including the declined and the unconverted, and nothing addresses retention, vendor personnel access under the hosted model, or the boundary between one insurer's data and anything the vendor computes across its base.

No security page, trust portal or enumerated certification located. One near miss recorded and deliberately not credited, because it illustrates the source test cleanly: a customer engagement story describes that client seeking an architecture with formalised security and confidentiality. That is the buyer's stated requirement, not a credential the vendor claims, and it earns nothing. Queued as a cheap check rather than treated as evidence of absence.

Graded C, and the reasoning is worth recording because this vendor is a clean illustration of a distinction that keeps recurring. The parent group genuinely holds regulatory permissions in many jurisdictions, as a licensed insurance and reinsurance broker. Those licences attach to broking, a different business, and confer nothing on a software product licensed to insurers who make their own filings in their own names.

A group licence for an adjacent activity is not standing for the graded product, exactly as an exchange listing is not. The contrast within this pocket is the sharp part: the smallest vendor here holds an A because regulators examined and approved its models, and the largest and most regulated group here holds a C because its models have never been through that process.

No fairness testing, protected characteristic handling or disparate impact analysis published, on a suite that includes a dedicated price optimisation component. That is the second vendor in this pocket selling price optimisation with no fairness position beside it, which turns a single observation into a pattern: optimisation against a customer's propensity to accept rather than their risk is the practice consumer fairness rules in this market were written to address, and neither vendor that sells it discloses testing.

The two vendors in the pocket that do disclose bias testing sell risk modelling rather than optimisation. Interpretability is credited under model risk and deliberately not counted again here, because being able to explain a model is not the same as testing who it disadvantages.

Nothing published on responsibility for a wrong outcome. The exposure follows directly from the scale claim: a rating engine serving billions of quotes a day distributes an error at that same rate, and nothing addresses detection, correction, who bears the cost of a mispriced book, or what happens to consumers already sold at the wrong price. Historic model retention means the state of the world at the time can be reconstructed, which is a forensic capability and answers a different question from who is responsible.

No model, family, provider or version named for the generative capability introduced in the current release, which is described only as advanced generative techniques applied to unstructured data in claims and underwriting. One adjacent property is recorded and deliberately not credited, because crediting it would stretch the rule: the environment lets customers deploy their own Python models alongside the vendor's, which is a statement about customer built models and not about what the vendor's own generative features depend on. The route that earns a B on this axis is publishing the architecture of the dependency, and that has not been done here.

Strong end of B on two specifics rather than on partner logos. First, a published performance envelope: the rate delivery component is stated to scale to billions of quotes a day with low latency and high resiliency for business critical decisions, which is the same class of artifact as naming a messaging standard or publishing a schema, and it is the second instance this session after Murex.

Second, the environment supports joint deployment of customer written Python models alongside the vendor's own in a single production path, with stated governance requirements for deploying open source code, which is a real interoperability commitment rather than an integration claim. Held at B: no policy administration connectors named, no interface documentation published.

Hosted delivery exists and was enhanced in the current release, and traditional licensing evidently continues, but neither is documented as an option set with what a buyer gets in each. No hosting regions, no residency commitment, no responsibility split, for software operating across 35 markets and processing quote and policy data subject to different national rules in each.

No pricing published, and the product structure makes the opacity larger than usual: the suite is modular and components are explicitly licensed in conjunction with one another, so a buyer cannot determine either a price or which combination of modules a given capability requires. Emblem is described as licensed with Radar Base and gradient boosting is described as available when licensed alongside another product, which tells a buyer that bundling matters and nothing about what it costs.

More than 1,000 client companies using the insurance software across six continents and 35 markets, stated by the vendor, covering personal and commercial lines and reaching pricing, underwriting, portfolio management, claims and reserving. The company claims most of the world's leading insurance groups among users, and the division is sized at over 1,700 staff. This is the historical incumbent in actuarial pricing software and the installed base reflects it.